A comparative study of kriging and deep learning methods for groundwater level estimation - A case study of Shenshan special cooperation zone
-
摘要: 掌握区域地下水位分布是地下水资源评价与环境保护的重要基础。由于区域尺度观测的地下水位数据有限,克里金插值与深度学习方法逐渐被用于区域地下水位预测,但两者的适用性及鲁棒性缺乏对比分析。针对这个问题,本文基于239口监测井水位,采用普通克里金、融合地表高程的协同克里金、深度学习方法估计深汕合作区地下水位空间分布,调查三种方法在区域地下水位预测中的应用潜力。结果显示,当训练样本量为76口水位数据时,考虑了地表高程信息的协同克里金法明显优于普通克里金法与深度学习法。当训练样本量增加到163口水位数据时,普通克里金、协同克里金及深度学习法的预测精度都明显提升,三种方法拟合验证数据集的RMSE相差很小,但不同方法之间预测水位的空间分布特征仍存在明显差异。结果表明,当观测数据稀疏时,融合高程信息的协同克里金的预测精度显著高于普通克里金和深度学习方法,而当观测数据密集时,三种方法预测精度接近。Abstract: The knowledge of regional groundwater level distribution is an important foundation for groundwater resource evaluation and environmental protection. Due to the limited groundwater level data observed at regional scale, kriging interpolation and deep learning methods are gradually used for regional groundwater level prediction, but their applicability and robustness lack comparative analysis. In this paper, spatial interpolation of groundwater levels in Shenshan special cooperation zone was carried out using ordinary kriging, cokriging and deep learning methods to explore the potential of the three methods in the practical application of regional groundwater level predictions. In order to investigate the effect of the training set sample size on the prediction effect of the three methods, 239 monitoring wells were divided into two groups of 76 and 163 wells for the training of the three models, respectively. The results showed that the cokriging, which considered surface elevations, was significantly better than the ordinary kriging and deep learning in regional groundwater levels predictions, when the training sample size was 76. When the training sample size was increased to 163, the prediction accuracies of ordinary kriging, cokriging, and deep learning were significantly improved. The RMSEs of the three methods on the validation dataset differed very little, but the spatial distribution characteristics of the predicted regional groundwater levels still differed significantly among the methods. In summary, when the observed data are sparse, the prediction accuracy of cokriging incorporating surface elevation is significantly higher than that of ordinary kriging and deep learning. However, the prediction accuracies of the three methods are close to each other, when the observed data are dense.
-
Key words:
- Groundwater level /
- Kriging method /
- Deep learning /
- Shenshan special cooperation zone.
点击查看大图
计量
- 文章访问数: 82
- PDF下载量: 8
- 被引次数: 0